A Playlist Generation System Using Annotations about Listeners' Preference and Situation

Katsuhiko KAJI
Graduate School of Information Science, Nagoya University

NTT Communication Science Laboratories
Katashi NAGAO
EcoTopia Science Institute, Nagoya University

1 Introduction

Recently, research on music recommendation and automatic playlist generation is actively in progress. These research is nessesary to present music suitable for each user from vast amount of music. Several researches on music recommendation have been reached the conclusion that collaborative filtering and use of musical metadata (genre, artist, etc.) is efficient to musical recommendation.

On the other hand, everal studies have been performed on annotation that associates the multimedia content with metadata.Annotation permits content's advanced use, such as semantic-content-based retrieval and simmarization.

We structured a playlist recommendation system using annotation about listener's interpretation.In addition to corrabolative filtering, it changes base playlist to the suitable one for the listener's preference by transcoding.Additionally, by system and listener's interactions, listener's preference is updated.

2 Presumption of musical similarity by using lyrics and annotations

2.1 Musical triple feature values

There are several feature values that is efficient to musical semantic analysis. For example, beat analyzed from audible signal and musical metadata such as artist and genre is useful for music recommendation. In this research, we adopt following three values as musical feature values: Lyrics, Musical Scene, and Listening situation. Beth Rogan proposed that it is able to musical semantic analysis by using lyrics.

As concerns musical scene and listening situation, they strongly depend on listener's aspect. For that reason, it is hard for automatic analysis to collect such information. Therefore such annotations have been collected for each song by musical questionnare system, and we applied these annotations as musical feature values.

2.2 Calculation of musical similarity

Figure shows vector spaces that each feature value is mapped.For preparation, keyword is extracted by TF*IDF from each lyrics.In terms of musical scene and listening situation, average of annotations is regard to feature value of each music.Cosine coefficient of any two songs is derived from the feature spaces.Then musical similarity can be calculated by each weighted cosine coefficient.

Triple feature spaces

Fugure1: Triple feature spaces

System holds listener's favorite song from history of playlist generation.Given that listener's ideal song is an average song of his favorite song,the user can be mapped to above-mentioned scene feature spaces of lyrics and musical.In the way of listener's situation, it can be mapped when system acquired listeners input of situation at the moment.Whereat similarity between arbitrary song and listener can be calculated.

There are several studies about music recommendation.They have concluded that it is effective to recommend music by using musical meta-data such as genre or artist.Also our system can import these effective metadata so that these feature value is mapped to multiple feature spaces.

3 Playlist recommendation system

3.1 Sequence of playlist generation

System Architecture

Fugure2: System Architecture

Figure shows sequence of playlist generation.The system generate playlist through these three phase.In first phase, initial playlist is found from playlist database by collaborative filtering.To find similar listener, listener's similarity can be calculated by cosine coefficient from triple feature spaces.We regard that similar listeners are over certain threshold. Then, similarity of the listener's situation and the situation in which each playlist is created is derived. From the playlist that is over cettain threshold, one initial playlist is searched in consideration of listeners' similarity and situation's similarity.

Second is transcoding phase. In this process, the initial playlist is imploved to be suitable for the listener's preference. In concrete terms, music that does not suit listener's preference is removed from the playlist. Besides, music that listener is not heard is involved to the playlist at the rate of 30%.

An example of generated playlist

Fugure3: An example of generated playlist

Through avobe-mentioned process, improved playlist is displayed like figure .Generated playlist is displayed at the left side and music player is embedded above of the playlist.Listener can operate the player as well as general music player.At the right side, attractive information related to the playlist is displayed.This information is described in detail at section 4.

Listener actually can enjoy the playlist and feedback some information to the system such aswhether the music is suitable for preference and situation.The system update listener's preference by using these feedback along with presenting fine-tuned playlist.

3.2 Acquirement of listening history and its reflection to user profile

Generally, as ideal information that is introduced into user profile, histry to have appreciated contents is given. To collect the information, the system embeds playlist player into web browser. Listener can operate the player like popular music player. The player feedbacks listener's operation history such as how many times was the music listened as needed. And then, user profile is updated with the information.

These operation history can not be useful for updating user preference, but useful as musical annotation. Annotation ,in the shape of listener's aspect, is not equally collected to each music. And so, music that is lack of annotation can compensate information by using operation history. For example, when a music is aboundingly listened in certain situation, it can be said that the music is suitable in the situation.

4 Future Work

Future work will contain method to enhance value of playlist such as liner-notes generation in accord with the playlist. Liner-notes can interest users to listening actually. Additionally, we are considering about highlight playlist generation by using musical annotation.